NVIDIA Unveils Ultrasound Raw2Insights-US

NV-Raw2Insights-US is an ultrasound reconstruction model unveiled by NVIDIA and Siemens Healthineers that trains directly on raw ultrasound channel data rather than completed images.

The first application is patient-specific speed of sound estimation. This corrects ultrasound focus and aims for real-time adaptive imaging by leveraging information that is otherwise simplified away and lost during conventional beamforming.

The deployment demo uses NVIDIA Holoscan and Holoscan Sensor Bridge (HSB) to read raw data from the DisplayPort output of an ACUSON Sequoia ultrasound scanner and transmit it to the GPU via Data over DisplayPort. The data is then aggregated on NVIDIA IGX and processed by inference on a Blackwell-class GPU, with the resulting speed-of-sound map sent back to the scanner to assist live image focusing.

The key points emphasized by the unveiled architecture are as follows.

• Software-only integration: Connects with existing medical devices in a software-centric way
• Software-defined ultrasound: Improves functionality through software updates
• Modular expansion: Once raw data is in GPU memory, other AI models can be easily added

Model weights, datasets, and a GitHub repository were also released, and the technology is explicitly stated to currently be at the investigational development stage, not cleared for commercial sale in the US and elsewhere.